Natural Scene Race Bib Number Detection

Using machine learning algorithms, including OpenCV, NIVIDIA's cuDNN, & Darknet's YOLOv4 to detect numbers on racing bibs found in natural image scene.

Overview and Background

Racing Bib Number (RBN) detection and recognition contain the interesting tasks of both finding the location of bib attached to a person in a natural scene & then inferring the text detection on the bib itself. To break these tasks down further, text recognition requires training steps, including finding the area of the bib on a person and then inferring the numbers on the bib. This project uses the research & experience from prior implementations to apply a working Convoltuional Neural Network (CNN) to detect race bib numbers in both images & video.

This repo investigates the use of Convolutional Neural Networks (CNN) and specifically, NVIDIA's cuDNN & Darknet's You Only Look Once ver. 4 (YOLOv4), to detect Racing Bib Numbers (RBNR) in a natural image scene. Leveraging publically available & labeled datasets from previous research (please see reference section below for addt'l information), I achieve a mean average precision (mAP) on the following:

• 96% mAP on Street View House Number dataset training
• 99% mAP on Race Bib Detection in Natural Scene dataset training

For future work, I would like to continue to train the model with datasets similar to the Street View House Numbers & more Race Bib Number images. Adjusting the configuration files to include more batches tweaks within the layers could be beneficial. Additionally, I would like to transfer the learning weights to edge compute hardware (i.e. NVIDIA's Jetson Nano or Raspberr Pi) to display a use case for edge computing with live streaming video capture.

Sample Video Output of Detection

YouTube Presentation

To support the submission of this project to UMBC's Data Science Program, class DATA690: Applied AI, here is the youtube containing presentation.

https://youtu.be/xfVfr0KmhYY

Data Details

• Street View House Numbers (SVHN) Dataset: This dataset was curated for the NIPS Workshop on Deep Learning and Unsupervised Feature Learning from 2011. For additional information regarding the dataset, please contact streetviewhousenumbers@gmail.com.
• Racing Bib Number Recognition Dataset: Supporting dataset for the published paper leveraging a different approach than we cover here.

Table of Contents

BibObjectDetection
|__ notebooks-utils-data
|   |__ 01 - Prepocessing & Training SVHN YOLOv4-tiny Darknet.ipynb  
|   |__ 02 - Digit Detection Validation Using RBNR Data.ipynb 
|   |__ 03 - Preprocessing Racing Bib Numbers (RBNR) Datasets.ipynb
|   |__ 04 - Run Yolov4 Tiny on RBNR Data.ipynb
|   |__ 05 - Bib Detection Validation & Demo.ipynb
|   |__ utils.py
|   |__ VIDEO0433.mp4
|   |__ output_marathon.mp4
|   |__ BibDetectorSample.jpeg
|__ presentation
|   |__ RaceBibDetection_Presentation.pdf
weights-classes
|   |__ SVHN_obj.names
|   |__ RBNR_obj.data 
|   |__ SVHN_custom-yolov4-tiny-detector.cfg
|   |__ SVHN_custom-yolov4-tiny-detector_best.weights
|   |__ RBNR_custom-yolov4-tiny-detector.cfg
|   |__ RBNR_custom-yolov4-tiny-detector_best.weights
README.md

References

• A. Apap and D. Seychell, “Marathon bib number recognition using deep learning,” in 2019 11th International Symposium on Image and Signal Processing and Analysis (ISPA), 2019, pp. 21–26.
• E. Ivarsson and R. M. Mueller, “Racing bib number recognition using deep learning,” 2019.
• P. Hernández-Carrascosa, A. Penate-Sanchez, J. Lorenzo-Navarro, D. Freire-Obregón, and M. Castrillón-Santana, “TGCRBNW: A Dataset for Runner Bib Number Detection (and Recognition) in the Wild,” in 2020 25th International Conference on Pattern Recognition (ICPR), 2021, pp. 9445–9451.
• G. Carty, M. A. Raja, and C. Ryan, “Running to Get Recognised,” in International Symposium on Signal Processing and Intelligent Recognition Systems, 2020, pp. 3–17.
• N. Boonsim, “Racing bib number localization on complex backgrounds,” WSEAS Trans. Syst. Control, vol. 13, pp. 226–231, 2018.
• RoboFlow: https://blog.roboflow.com/train-yolov4-tiny-on-custom-data-lighting-fast-detection/
• OpenCV: https://opencv.org/
• NVIDIA cuDNN: https://developer.nvidia.com/cudnn
• Eric Bayless: https://github.com/ericBayless/bib-detector

Project Curation

Note also that notebooks were created into Google Collaboratory to take advantage of GPU throughput to speed the training of the CNN.

Contributors : Lee Whieldon

Languages    : Python
Tools/IDE    : Google Colab, Darknet, cuDNN
Libraries    : h5py, numpy, cv2, os, matplotlib.pyplot, scipy.io, pandas, imgaug

Assignment Submitted     : May 2022